Co-evolution of Language and of the Language Acquisition Device Ted Briscoe ejb¢cl, cam. ac. uk Computer Laboratory University of Cambridge Pembroke Street Cambridge CB2 3QG, UK Abstract A new account of parameter setting dur- ing grammatical acquisition is presented in terms of Generalized Categorial Grammar embedded in a default inheritance hierar- chy, providing a natural partial ordering on the setting of parameters. Experiments show that several experimentally effective learners can be defined in this framework. Ew)lutionary simulations suggest that a lea.rner with default initial settings for pa- rameters will emerge, provided that learn- ing is memory limited and the environment of linguistic adaptation contains an appro- priate language. 1 Theoretical Background Grmnnmtical acquisition proceeds on the basis of a partial genotypic specifica.tion of (universal) grmn- mar (UG) complemented with a learning procedure elmbling the child to complete this specification ap- propriately. The parameter setting frainework of Chomsky (1981) claims that learning involves fix- ing the wdues of a finite set of finite-valued param- eters to select a single fully-specified grammar from within the space defined by the genotypic specifi- cation of UG. Formal accounts of parameter set- ting have been developed for small fragments but even in these search spaces contain local maxima and subset-superset relations which may cause a learner to converge to an incorrect grammar (Clark, 1992; Gibson and Wexler, 1994; Niyogi and Berwick, 1995). The solution to these problems involves defin- ing d(,fault, umnarked initial values for (some) pa- rameters and/or ordering the setting of paraineters during learning. Bickerton (1984) argues for the Bioprograin Hy- pothesis a.s an explanation for universal similarities between historically unrelated creoles, and for the rapid increase in gramlnatical complexity accompa- nying the transition from pidgin to creole languages. Prom the perspective of the parameters framework, the Bioprogram Hypothesis claims that children are endowed genetically with a UG which, by default, specifies the stereotypical core creole grammar, with right-branching syntax and subject-verb-object or- der, as in Saramaccan. Others working within the parameters framework have proposed unmarked, de- fault parameters (e.g. Lightfoot, 1991), but the Bio- program Hypothesis can be interpreted as towards one end of a continuum of proposals ranging from all parameters initially unset to all set to default values. 2 The Language Acquisition Device A model of the Language Acquisition Device (LAD) incorporates a UG with associated parameters, a parser, and an algorithm for updating initial param- eter settings on parse failure during learning. 2.1 The Grammar (set) Basic categorial grammar (CG) uses one rule of ap- plication which combines a functor category (con- taining a slash) with an argument category to form a derived category (with one less slashed argument category). Grammatical constraints of order and agreement are captured by only allowing directed application to adjacent matching categories. Gener- alized Categorial Grammar (GCG) extends CG with further rule schemata) The rules of FA, BA, gen- eralized weak permutation (P) and backward and forward colnposition (I?C, BC) are given in Fig- ure 1 (where X, Y and Z are category variables, [ is a vm'iable over slash and backslash, and denotes zero or more further flmctor arguments). Once pernmtation is included, several semantically l\¥ood (1993) is a general introduction to Categorial Grammar mid extensions to the basic theory. The most closely related theories to that presented here are those of Steedman (e.g. 1988) and Hoffman (1995). 418 X/Y Y ~ X Y X\Y ~ X Forward Application: A y [X(y)] (y) ::~ X(y) Backward Application: A y [X(y)] (y) =~ X(y) X/Y Y/Z ~ X/Z Y\Z X\Y ~ X\Z Forward Composition: y [X(y)] A z [Y(z)] =~ A z [X(Y(z))] Backward Composition: z [Y(z)] A y [X(y)] ~ A z [X(Y(z))] (Generalized Weak) Permutation: (XIY1) IY, ~ (XIYn)IYI A Yn ,Yl [X(yl ,y,.)] =V A Yl,Y [X(yl ,Yn)] Figure 1: GCG Rule Schemata Kim loves Sandy NP (S\NP)/NP NP kim' A y,x [love'(x y)] sandy' P (S/NP)\NP A x,y [love'(x y)] -BA S/NP A y [love'(kim' y)] FA S love'(kim' sandy') Figure 2: GCG Derivation for Kim loves Sandy equivalent derivations for Kim loves Sandy become available, Figure 2 shows the non-conventional left- branching one. Composition also allows alterna- tive non-conventional semantically equivalent (left- branching) derivations. GCG as presented is inadequate as an account of UG or of any individual grammar. In particular, the definition of atomic categories needs extending to deal with featural variation (e.g. Bouma and van Noord, 1994), and the rule schemata, especially com- position and weak permutation, must be restricted in various parametric ways so that overgeneration is prevented for specific languages. Nevertheless, GCG does represent a plausible kernel of UG; Hoff- man (1995, 1996) explores the descriptive power of a very similar system, in which generalized weak per- mutation is not required because functor arguments are interpreted as multisets. She demonstrates that this system can handle (long-distance) scrambling elegantly and generates mildly context-sensitive lan- guages (Joshi et al, 1991). The relationship between GCG as a theory of UG (GCUG) and as a the specification of a particu- lar grammar is captured by embedding the theory in a default inheritance hierarchy. This is repre- sented as a lattice of typed default feature structures (TDFSs) representing subsumption and default in- heritance relationships (Lascarides et al, 1996; Las- carides and Copestake, 1996). The lattice defines intensionally the set of possible categories and rule schemata via type declarations on nodes. For ex- ample, an intransitive verb might be treated as a subtype of verb, inheriting subject directionality by default from a type gendir (for general direction). For English, gendir is default right but the node of the (intransitive) functor category, where the direc- tionality of subject arguments is specified, overrides this to left, reflecting the fact that English is pre- dominantly right-branching, though subjects appear to the left of the verb. A transitive verb would in- herit structure from the type for intransitive verbs and an extra NP argument with default directional- ity specified by gendir, and so forth. 2 For the purposes of the evolutionary simulation described in §3, GC(U)Gs are represented as a se- quence of p-settings (where p denotes principles or parameters) based on a flat (ternary) sequential en- coding of such default inheritance lattices. The in- 2Bouma and van Noord (1994) and others demon- strate that CGs can be embedded in a constraint-based representation. Briscoe (1997a,b) gives further details of the encoding of GCG in TDFSs. 419 NP N S gen-dir subj-dir applic AT AT AT DR DL DT NP gendir applic S N subj-dir AT DR DT AT AT DL "applic NP N gen-dir subj-dir S DT AT AT DR DL AT Figure 3: Sequential encodings of the grammar fragment heritance hierarchy provides a partial ordering on parameters, which is exploited in the learning pro- cedure. For example, the atomic categories, N, NP and S are each represented by a parameter en- coding the presence/absence or lack of specification (T/F/?) of the category in the (U)G. Since they will be unordered in the lattice their ordering in the se- quential coding is arbitrary. However, the ordering of the directional types gendir and subjdir (with values L/R) is significant as the latter is a more spe- cific type. The distinctions between absolute, de- fault or unset specifications also form part of the encoding (A/D/?). Figure 3 shows several equiva- lent and equally correct sequential encodings of the fragment of the English type system outlined above. A set of grammars based on typological distinc- tions defined by basic constituent order (e.g. Green- berg, 1966; Hawkins, 1994) was constructed as a (partial) GCUG with independently varying binary- valued parameters. The eight basic language fami- lies are defined in terms of the unmarked order of verb (V), subject (S) and objects (0) in clauses. Languages within families further specify the order of modifiers and specifiers in phrases, the order of ad- positions and further phrasal-level ordering param- eters. Figure 4 list the language-specific ordering parameters used to define the full set of grammars in (partial) order of generality, and gives examples of settings based on familiar languages such as "En- glish", "German" and "Japanese". 3 "English" de- fines an SVO language, with prepositions in which specifiers, complementizers and some modifiers pre- cede heads of phrases. There are other grammars in the SVO family in which all modifers follow heads, there are postpositions, and so forth. Not all combi- nations of parameter settings correspond to attested languages and one entire language family (OVS) is unattested. "Japanese" is an SOV language with 3Throughout double quotes around language names are used as convenient mnemonics for familiar combina- tions of parameters. Since not all aspects of these actual languages are represented in the grammars, conclusions about actual languages must be made with care. postpositions in which specifiers and modifiers follow heads. There are other languages in the SOV family with less consistent left-branching syntax in which specifiers and/or modifiers precede phrasal heads, some of which are attested. "German" is a more complex SOV language in which the parameter verb- second (v2) ensures that the surface order in main clauses is usually SVO. 4 There are 20 p-settings which determine the rule schemata available, the atomic category set, and so forth. In all, this CGUG defines just under 300 grammars. Not all of the resulting languages are (stringset) distinct and some are proper subsets of other languages. "English" without the rule of per- mutation results in a stringset-identical language, but the grammar assigns different derivations to some strings, though the associated logical forms are identical. "English" without composition results in a subset language. Some combinations of p-settings result in 'impossible' grammars (or UGs). Others yield equivalent grammars, for example, different combinations of default settings (for types and their subtypes) can define an identical category set. The grammars defined generate (usually infinite) stringsets of lexical syntactic categories. These strings are sentence types since each is equivalent to a finite set of grammatical sentences formed by selecting a lexical instance of each lexicai category. Languages are represented as a finite subset of sen- tence types generated by the associated grammar. These represent a sample of degree-1 learning trig- gers for the language (e.g. Lightfoot, 1991). Subset languages are represented by 3-9 sentence types and 'full' languages by 12 sentence types. The construc- tions exemplified by each sentence type and their length are equivalent across all the languages defined by the grammar set, but the sequences of lexical cat- egories can differ. For example, two SOV language renditions of The man who Bill likes gave Fred a 4Representation of the vl/v2 parameter(s) in terms of a type constraint determining allowable functor cate- gories is discussed in more detail in Briscoe (1997b). 420 gen vl n subj obj v2 mod spec relcl adpos compl Engl R F R L R F R R R R R Ger R F R L L T R R R R R Jap L F L L L F L L L L ? Figure 4: The Grammar Set - Ordering Parameters present, one with premodifying and the other post- modifying relative clauses, both with a relative pro- noun at the right boundary of the relative clause, are shown below with the differing category highlighted. Bill likes who the-man a-present Fred gave NP8 (S\NP,)\NPo Rc\(S\NPo) NPs\Rc NPo2 NPol ((S\NPs)\NPo2)\NPol The-man Bill likes who a-present Fred gave NPs/Rc NPs (S\NPs)\NPo Rc\(S\NPo) NPo2 NPol ((S\NPs)\NPo2)\NPol 2.2 The Parser The parser is a deterministic, bounded-context stack-based shift-reduce algorithm. The parser op- erates on two data structures, an input buffer or queue, and a stack or push down store. The algo- rithm for the parser working with a GCG which in- cludes application, composition and permutation is given in Figure 5. This algorithm finds the most left- branching derivation for a sentence type because Re- duce is ordered before Shift. The category sequences representing the sentence types in the data for the entire language set are designed to be unambiguous relative to thi s 'greedy', deterministic algorithm, so it will always assign the appropriate logical form to each sentence type. However, there are frequently al- ternative less left-branching derivations of the same logical form. The parser is augmented with an algorithm which computes working memory load during an analy- sis (e.g. Baddeley, 1992). Limitations of working memory are modelled in the parser by associating a cost with each stack cell occupied during each step of a derivation, and recency and depth of process- ing effects are modelled by resetting this cost each time a reduction occurs: the working memory load (WML) algorithm is given in Figure 6. Figure 7 gives the right-branching derivation for Kim loves Sandy, found by the parser utilising a grammar without per- mutation. The WML at each step is shown for this derivation. The overall WML (16) is higher than for the left-branching derivation (9). The WML algorithm ranks sentence types, and 1. The Reduce Step: if the top 2 cells of the stack are occupied, then try a) Application, if match, then apply and goto 1), else b), b) Combination, if match then apply and goto 1), else c), c) Permutation, if match then apply and goto 1), else goto 2) 2. The Shift Step: if the first cell of the Input Buffer is occupied, then pop it and move it onto the Stack to- gether with its associated lexical syntactic cat- egory and goto 1), else goto 3) 3. The Halt Step: if only the top cell of the Stack is occupied by a constituent of category S, then return Success, else return Fail The Match and Apply operation: if a binary rule schema matches the categories of the top 2 cells of the Stack, then they are popped from the Stack and the new category formed by applying the rule schema is pushed onto the Stack. The Permutation operation: each time step lc) is visited during the Reduce step, permutation is ap- plied to one of the categories in the top 2 cells of the Stack until all possible permutations of the 2 cate- gories have been tried using the binary rules. The number of possible permutation operations is finite and bounded by the maximum number of arguments of any functor category in the grammar. Figure 5: The Parsing Algorithm 421 Stack Input Buffer Operation Step WML Kim loves Sandy 0 0 Kim:NP:kim ~ loves Sandy Shift 1 1 loves:(S\NP)/NP:A y,x(love' x, y) Sandy Shift 2 3 Kim:NP:kim ~ Sandy:NP:sandy ~ Shift 3 6 loves:(S\NP)/NP:A y,x(love' x, y) Kim:NP:kim ~ loves Sandy:S/NP:A x(love' x, sandy') Reduce (A) 4 Kim:NP:kim ~ Kim loves Sandy:S:(love' kim ~, sandy ~) Reduce (A) 5 Figure 7: WML for Kim loves Sandy After each parse step (Shift, Reduce, Halt (see Fig 5): 1. Assign any new Stack entry in the top cell (in- troduced by Shift or Reduce) a WML value of 0 2. Increment every Stack cell's WML value by 1 3. Push the sum of the WML values of each Stack cell onto the WML-record When the parser halts, return the sum of the WML- record gives the total WML for a derivation Figure 6: The WML Algorithm thus indirectly languages, by parsing each sentence type from the exemplifying data with the associ- ated grammar and then taking the mean of the WML obtained for these sentence types. "En- glish" with Permutation has a lower mean WML than "English" without Permutation, though they are stringset-identical, whilst a hypothetical mix- ture of "Japanese" SOV clausal order with "En- glish" phrasal syntax has a mean WML which is 25% worse than that for "English". The WML algorithm is in accord with existing (psycholinguistically- motivated) theories of parsing complexity (e.g. Gib- son, 1991; Hawkins, 1994; Rambow and Joshi, 1994). 2.3 The Parameter Setting Algorithm The parameter setting algorithm is an extension of Gibson and Wexler's (1994) Trigger Learning Al- gorithm (TLA) to take account of the inheritance- based partial ordering and the role of memory in learning. The TLA is error-driven - parameter set- tings are altered in constrained ways when a learner cannot parse trigger input. Trigger input is de- fined as primary linguistic data which, because of its structure or context of use, is determinately un- parsable with the correct interpretation (e.g. Light- foot, 1991). In this model, the issue of ambigu- ity and triggers does not arise because all sentence types are treated as triggers represented by p-setting schemata. The TLA is memoryless in the sense that a history of parameter (re)settings is not maintained, in principle, allowing the learner to revisit previous hypotheses. This is what allows Niyogi and Berwick (1995) to formalize parameter setting as a Markov process. However, as Brent (1996) argues, the psy- chological plausibility of this algorithm is doubt- ful - there is no evidence that children (randomly) move between neighbouring grammars along paths that revisit previous hypotheses. Therefore, each parameter can only be reset once during the learn- ing process. Each step for a learner can be defined in terms of three functions: P-SETTING, GRAMMAR and PARSER, as: PARSERi(GRAMMAR/(P-SETTING/(Sentence j))) A p-setting defines a grammar which in turn defines a parser (where the subscripts indicate theoutput of each function given the previous trigger). A param- eter is updated on parse failure and, if this results in a parse, the new setting is retained. The algo- rithm is summarized in Figure 8. Working mem- ory grows through childhood (e.g. Baddeley, 1992), and this may assist learning by ensuring that trigger sentences gradually increase in complexity through the acquisition period (e.g. Elman, 1993) by forcing the learner to ignore more complex potential triggers that occur early in the learning process. The WML of a sentence type can be used to determine whether it can function as a trigger at a particular stage in learning. 422 Data: {$1, S2, Sn} unleSs PARSERi( GRAMMARi(P-SETTINGi(Sj ) ) ) : Success then p-settingj = UPDATE(p-settings) unless PARSERj (GRAMMARj (P-SETTINGj (Sj))) Success then RETURN p-settings/ else RETURN p-settingsy Update: Reset the first (most general) default or unset pa- rameter in a left-to-right search of the p-set accord- ing to the following table: Input: D 1 D0 ? ? ] Output: R 0 R 1 ? 1/0 (random) I (where 1 = T/L and 0 = F/R) Figure 8: The Learning Algorithm 3 The Simulation Model The computational simulation supports the evolu- tion of a population of Language Agents (LAgts), similar to Holland's (1993) Echo agents. LAgts gen- erate and parse sentences compatible with their cur- rent p-setting. They participate in linguistic inter- actions which are successful if their p-settings are compatible. The relative fitness of a LAgt is a func- tion of the proportion of its linguistic interactions which have been successful, the expressivity of the language(s) spoken, and, optionally, of the mean WML for parsing during a cycle of interactions. An interaction cycle consists of a prespecified number of individual random interactions between LAgts, with generating and parsing agents also selected ran- domly. LAgts which have a history of mutually suc- cessful interaction and high fitness can 'reproduce'. A LAgt can 'live' for up to ten interaction cycles, but may 'die' earlier if its fitness is relatively low. It is possible for a population to become extinct (for example, if all the initial LAgts go through ten in- teraction cycles without any successful interaction occurring), and successful populations tend to grow at a modest rate (to ensure a reasonable proportion of adult speakers is always present). LAgts learn during a critical period from ages 1-3 and reproduce from 4-10, parsing and/or generating any language learnt throughout their life. During learning a LAgt can reset genuine param- Variables Typical Values Population Size 32 Interaction Cycle 2K Interactions Simulation Run 50 Cycles Crossover Probability 0.9 Mutation Probability 0 Learning memory limited yes critical period yes Figure 9: The Simulation Options (Cost/Benefits per sentence (1-6); summed for each LAgt at end of an interaction cycle and used to cal- culate fitness functions (7-8)): 1. Generate cost: 1 (GC) 2. Parse cost: ! (PC) 3. Generate subset language cost: 1 (GSC) 4. Parse failure cost: 1 (PF) 5. Parse memory cost: WML(st) 6. Interaction success benefit: 1 (SI) 7. Fitness(WML): SI GC • GC+PC X GC+GSC X 8. Fitness(-~WML): sI cc GC+PC X CC [-GSC Figure 10: Fitness Functions eters which either were unset or had default settings 'at birth'. However, p-settings with an absolute value (principles) cannot be altered during the life- time of an LAgt. Successful LAgts reproduce at the end of interaction cycles by one-point crossover of (and, optionally, single point mutation of) their ini- tial p-settings, ensuring neo-Darwinian rather than Lamarckian inheritance. The encoding of p-settings allows the deterministic recovery of the initial set- ting. Fitness-based reproduction ensures that suc- cessful and somewhat compatible p-settings are pre- served in the population and randomly sampled in the search for better versions of universal grammar, including better initial settings of genuine parame- ters. Thus, although the learning algorithm per se is fixed, a range of alternative learning procedures can be explored based on the definition of the inital set of parameters and their initial settings. Figure 9 summarizes crucial options in the simulation giving the values used in the experiments reported in §4 and Figure 10 shows the fitness functions. 423 4 Experimental Results 4.1 Effectiveness of Learning Procedures Two learning procedures were predefined - a default learner and an unset learner. These LAgts were ini- tialized with p-settings consistent with a minimal in- herited CGUG consisting of application with NP and S atomic categories. All the remaining p-settings were genuine parameters for both learners. The un- set learner was initialized with all unset, whilst the default learner had default settings for the parame- ters gendir and subjdir and argorder which spec- ify a minimal SVO right-branching grammar, as well as default (off) settings for comp and perm which determine the availability of Composition and Per- mutation, respectively. The unset learner represents a 'pure' principles-and-parameters learner. The de- fault learner is modelled on Bickerton's bioprogram learner. Each learner was tested against an adult LAgt initialized to generate one of seven full lan- guages in the set which are close to an at- tested language; namely, "English" (SVO, predom- inantly right-branching), "Welsh" (SVOvl, mixed order), "Malagasy" (VOS, right-branching), "Taga- log" (VSO, right-branching), "Japanese" (SOV, left-branching), "German" (SOVv2, predominantly right-branching), "Hixkaryana" (OVS, mixed or- der), and an unattested full OSV language with left- branching syntax. In these tests, a single learner in- teracted with a single adult. After every ten interac- tions, in which the adult randomly generated a sen- tence type and the learner attempted to parse and learn from it, the state of the learner's p-settings was examined to determine whether the learner had con- verged on the same grammar as the adult. Table 1 shows the number of such interaction cycles (i.e. the number of input sentences to within ten) required by each type of learner to converge on each of the eight languages. These figures are each calculated from 100 trials to a 1% error rate; they suggest that, in general, the default learner is more effective than the unset learner. However, for the OVS language (OVS languages represent 1.24% of the world's lan- guages, Tomlin, 1986), and for the unattested OSV language, the default (SVO) learner is less effective. So, there are at least two learning procedures in the space defined by the model which can converge with some presentation orders on some of the grammars in this set. Stronger conclusions require either ex- haustive experimentation or theoretical analysis of the model of the type undertaken by Gibson and Wexler (1994) and Niyogi and Berwick (1995). Unset Default None WML 15 39 26 -~WML 34 17 29 Table 2: Overall preferences for parameter types 4.2 Evolution of Learning Procedures In order to test the preference for default versus un- set parameters under different conditions, the five parameters which define the difference between the two learning procedures were tracked through an other series of 50 cycle runs initialized with either 16 default learning adult speakers and 16 unset learning adult speakers, with or without memory-limitations during learning and parsing, speaking one of the eight languages described above. Each condition was run ten times. In the memory limited runs, default parameters came to dominate some but not all pop- ulations. In a few runs all unset parameters dis- appeared altogether. In all runs with populations initialized to speak "English" (SVO) or "Malagasy" (VOS) the preference for default settings was 100%. In 8 runs with "Tagalog" (VSO) the same preference emerged, in one there was a preference for unset pa- rameters and in the other no clear preference. How- ever, for the remaining five languages there was no strong preference. The results for the runs without memory limita- tions are different, with an increased preference for unset parameters across all languages but no clear 100% preference for any individual language. Ta- ble 2 shows the pattern of preferences which emerged across 160 runs and how this was affected by the presence or absence of memory limitations. To test whether it was memory limitations during learning or during parsing which were affecting the results, another series of runs for "English" was per- formed with either memory limitations during learn- ing but not parsing enabled, or vice versa. Memory limitations during learning are creating the bulk of the preference for a default learner, though there appears to be an additive effect. In seven of the ten runs with memory limitations only in learning, a clear preference for default learners emerged. In five of the runs with memory limitations only in parsing there appeared to be a slight preference for defaults emerging. Default learners may have a fitness ad- vantage when the number of interactions required to learn successfully is greater because they will tend to converge faster, at least to a subset language. This will tend to increase their fitness over unset learners who do not speak any language until further into the 424 Learner Language SVO SVOvl VOS VSO SOV SOVv2 OVS OSV Unset 60 80 70 80 70 70 70 70 Default 60 60 60 60 60 60 80 70 Table 1: Effectiveness of Two Learning Procedures learning period. The precise linguistic environment of adaptation determines the initial values of default parameters which evolve. For example, in the runs initialized with 16 unset learning "Malagasy" VOS adults and 16 default (SVO) learning VOS adults, the learn- ing procedure which dominated the population was a variant VOS default learner in which the value for subjdir was reversed to reflect the position of the subject in this language. In some of these runs, the entire population evolved a default sub- jdir 'right' setting, though some LAgts always re- tained unset settings for the other two ordering pa- rameters, gendir and argo, as is illustrated in Fig- ure 11. This suggests that if the human language fac- ulty has evolved to be a right-branching SVO default learner, then the environment of linguistic adapta- tion must have contained a dominant language fully compatible with this (minimal) grammar. 4.3 Emergence of Language and Learners To explore the emergence and persistence of struc- tured language, and consequently the emergence of effective learners, (pseudo) random initialization was used. A series of simulation runs of 500 cycles were performed with random initialization of 32 LAgts' p-settings for any combination of p-setting values, with a probability of 0.25 that a setting would be an absolute principle, and 0.75 a parameter with unbi- ased allocation for default or unset parameters and for values of all settings. All LAgts were initialized to be age 1 with a critical period of 3 interaction cycles of 2000 random interactions for learning, a maximum age of 10, and the ability to reproduce by crossover (0.9 probability) and mutation (0.01 prob- ability) from 4-10. In around 5% of the runs, lan- guage(s) emerged and persisted to the end of the run. Languages with close to optimal WML scores typi- cally came to dominate the population quite rapidly. However, sometimes sub-optimal languages were ini- tially selected and occasionally these persisted de- spite the later appearance of a more optimal lan- guage, but with few speakers. Typically, a minimal subset language dominated - although full and inter- mediate languages did appear briefly, they did not survive against less expressive subset languages with a lower mean WML. Figure 12 is a typical plot of the emergence (and extinction) of languages in one of these runs. In this run, around 10 of the initial population converged on a minimal OVS language and 3 others on a VOS language. The latter is more optimal with respect to WML and both are of equal expressivity so, as expected, the VOS language ac- quired more speakers over the next few cycles. A few speakers also converged on VOS-N, a more expres- sive but higher WML extension of VSO-N-GWP- COMP. However, neither this nor the OVS language survived beyond cycle 14. Instead a VSO language emerged at cycle 10, which has the same minimal expressivity of the VOS language but a lower WML (by virtue of placing the subject before the object) and this language dominated rapidly and eclipsed all others by cycle 40. In all these runs, the population settled on sub- set languages of low expressivity, whilst the percent- age of absolute principles and default parameters in- creased relative to that of unset parameters (mean % change from beginning to end of runs: +4.7, +1.5 and -6.2, respectively). So a second identical set of ten was undertaken, except that the initial popula- tion now contained two SOV-V2 "German" speak- ing unset learner LAgts. In seven of these runs, the population fixed on a full SOV-V2 language, in two on the intermediate subset language SOV-V2-N, and in one on the minimal subset language SOV-V2-N- GWP-COMP. These runs suggest that if a full lan- guage defines the environment of adaptation then a population of randomly initialized LAgts is more likely to converge on a (related) full language. Thus, although the simulation does not model the devel- opment of expressivity well, it does appear that it can model the emergence of effective learning pro- cedures for (some) full languages. The pattern of language emergence and extinction followed that of the previous series of runs: lower mean WML lan- guages were selected from those that emerged during the run. However, often the initial optimal SVO-V2 itself was lost before enough LAgts evolved capable of learning this language. In these runs, changes in the percentages of absolute, default or unset p- settings in the population show a marked difference: 425 100 / 80 -"': / i 60 '',,": !'/ 40 V 2O 0 i 0 I0 I ; i i : ,,./"'_ .,-', ,,' " ' ,, -, ,,' ,: 'I/ "G0g"~ ~di~" "G0argo" - "G0subjdir f ,,v, j i / " '~'v i , ,/ i\},V I i ~ a I \q9 ,f I I I I 20 30 40 50 60 70 Interaction Cycles Q. q) "5 Figure 11: Percentage of each default ordering pa- rameter 45 40 35 30 25 20 15 10 5 0 i ; i i IL i "aa-S¢" "GB-OVS-N-P-C k "ge-y~,o-N ~ ,., ~GS-,VOS-N'., . ""GB-VOS-N-~WI~-COMP" k-:::." "G,8-VSOrN:GWP-COMP" - 'l ! t i-/~ i i ; i ! zi !'i ! z'! /~11 \' i z i V-"" " i ~ L /\ I '-V"'~':'( "'''', i I \ i 5 10 15 20 25 30 Interaction Cycles I /-'x, I I 35 40 45 50 Figure 12: Emergence of language(s) the mean number of absolute principles declined by 6.1% and unset parameters by 17.8%, so the num- ber of default parameters rose by 23.9% on average between the beginning and end of the 10 runs. This may reflect the more complex linguistic environment in which (incorrect) absolute settings are more likely to handicap, rather than simply be irrelevant to, the performance of the LAgt. 5 Conclusions Partially ordering the updating of parameters can result in (experimentally) effective learners with a more complex parameter system than that studied previously. Experimental comparison of the default (SVO) learner and the unset learner suggests that the default learner is more efficient on typologically more common constituent orders. Evolutionary sim- ulation predicts that a learner with default param- eters is likely to emerge, though this is dependent both on the type of language spoken and the pres- ence of memory limitations during learning and pars- ing. Moreover, a SVO bioprogram learner is only likely to evolve if the environment contains a domi- nant SVO language. The evolution of a bioprogram learner is a man- ifestation of the Baldwin Effect (Baldwin, 1896) - genetic assimilation of aspects of the linguistic envi- ronment during the period of evolutionary adapta- tion of the language learning procedure. In the case of grammar learning this is a co-evolutionary process in which languages (and their associated grammars) are also undergoing selection. The WML account of parsing complexity predicts that a right-branching SVO language would be a near optimal selection at a stage in grammatical development when complex rules of reordering such as extraposition, scrambling or mixed order strategies such as vl and v2 had not evolved. Briscoe (1997a) reports further exper- iments which demonstrate language selection in the model. Though, simulation can expose likely evolution- ary pathways under varying conditions, these might have been blocked by accidental factors, such as ge- netic drift or bottlenecks, causing premature fixa- tion of alleles in the genotype (roughly correspond- ing to certain p-setting values). The value of the simulation is to, firstly, show that a bioprogram learner could have emerged via adaptation, and sec- ondly, to clarify experimentally the precise condi- tions required for its emergence. Since in many cases these conditions will include the presence of constraints (working memory limitations, expressiv- ity, the learning algorithm etc.) which will remain causally manifest, further testing of any conclusions drawn must concentrate on demonstrating the ac- 426 curacy of the assumptions made about such con- straints. Briscoe (1997b) evaluates the psychological plausibility of the account of parsing and working memory. References Baddeley, A. (1992) 'Working Memory: the interface between memory and cognition', J. of Cognitive Neuroscience, vol.4.3, 281-288. Baldwin, J.M. (1896) 'A new factor in evolution', American Naturalist, vol.30, 441-451. Bickerton, D. (1984) 'The language bioprogram hy- pothesis', The Behavioral and Brain Sciences, vol. 7.2, 173-222. Bouma, G. and van Noord, G (1994) 'Constraint- based categorial grammar', Proceedings of the 32nd Assoc. for Computational Linguistics, Las Cruces, NM, pp. 147-154. Brent, M. (1996) 'Advances in the computational study of language acquisition', Cognition, vol. 61, 1-38. Briscoe, E.J. (1997a, submitted) 'Language Acquisi- tion: the Bioprogram Hypothesis and the Bald- win Effect', Language, Briscoe, E.J. (1997b, in prep.) Working memory and its influence on the development of human lan- guages and the human language faculty, Univer- sity of Cambridge, Computer Laboratory, m.s Chomsky, N. (1981) Government and Binding, Foris, Dordrecht. Clark, R. (1992) 'The selection of syntactic knowl- edge', Language Acquisition, vol.2.2, 83-149. Elman, J. (1993) 'Learning and development in neu- ral networks: the importance of starting small', Cognition, vol.48, 71-99. Gibson, E. (1991) A Copmutational Theory of Hu- man Linguistic Processing: Memory Limitations and Processing Breakdown, Doctoral disserta- tion, Carnegie Mellon University. Gibson, E. and Wexler, K. (1994) 'Triggers', Lin- guistic Inquiry, vol.25.3, 407-454. Greenberg, J. (1966) 'Some universals of grammar with particular reference to the order of mean- ingflll elements' in J. Greenberg (ed.), Univer- sals of Grammar, MIT Press, Cambridge, Ma., pp. 73-113. Hawkins, J.A. (1994) A Performance Theory of Order and Constituency, Cambridge University Press, Cambridge. Hoffman, B. (1995) The Computational Analysis of the Syntax and Interpretation of 'Free' Word Or- der in Turkish, PhD dissertation, University of Pennsylvania. Hoffman, B. (1996) 'The formal properties of syn- chronous CCGs', Proceedings o] the ESSLLI For- mal Grammar Conference, Prague. Holland, J.H. (1993) Echoing emergence: objectives, rough definitions and speculations for echo-class models, Santa Fe Institute, Technical Report 93- 04-023. Joshi, A., Vijay-Shanker, K. and Weir, D. (1991) 'The convergence of mildly context-sensitive grammar formalisms' in Sells, P., Shieber, S. and Wasow, T. (ed.), Foundational Issues in Natural Language Processing, MIT Press, pp. 31-82. Lascarides, A., Briscoe E.J. , Copestake A.A and Asher, N. (1995) 'Order-independent and persis- tent default unification', Linguistics and Philos- ophy, vo1.19.1, 1-89. Lascarides, A. and Copestake A.A. (1996, submit- ted) 'Order-independent typed default unifica- tion', Computational Linguistics, Lightfoot, D. (1991) How to Set Parameters: Argu- ments from language Change, MIT Press, Cam- bridge, Ma Niyogi, P. and Berwick, R.C. (1995) 'A markov language learning model for finite parameter spaces', Proceedings of the 33rd Annual Meet- ing of the Association for Computational Lin- guistics, MIT, Cambridge, Ma Rambow, O. and Joshi, A. (1994) 'A processing model of free word order languages' in C. Clifton, L. Frazier and K. Rayner (ed.), Perspectives on Sentence Processing, Lawrence Erlbaum, Hills- dale, NJ., pp. 267-301. Steedman, M. (1988) 'Combinators and grammars' in R. Oehrle, E. Bach and D. Wheeler (ed.), Cat- egorial Grammars and Natural Language Struc- tures, Reidel, Dordrecht, pp. 417-442. Tomlin, R. (1986) Basic Word Order: Functional Principles, Routledge, London. Wood, M.M. (1993) Categorial-Grammars, Rout- ledge, London. 427 . of the emergence (and extinction) of languages in one of these runs. In this run, around 10 of the initial population converged on a minimal OVS language and 3 others on a VOS language. The. schema matches the categories of the top 2 cells of the Stack, then they are popped from the Stack and the new category formed by applying the rule schema is pushed onto the Stack. The Permutation. virtue of placing the subject before the object) and this language dominated rapidly and eclipsed all others by cycle 40. In all these runs, the population settled on sub- set languages of low